OctoInform Beta

OpenAI preprint, September 2026 · §3. Ordered arrays and conditional locality · §3

Proposition 3.5 [Square locality]: Conditional laws of rows given the parameter array, of columns given parent rows, and of single cells and blocks given other columns

The square of a saturated array object is conditionally local.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Take a finite index set II and a standard Borel space YY of states, each state carrying a parameter. A random array of order two is a square matrix y=(ya,j)a,j∈Iy=(y_{a,j})_{a,j\in I} of states; an array of order n+2n+2 is the same square with cells ya,jy_{a,j} that are themselves arrays of order nn. An array object prescribes the laws of such arrays at every order.

The question is how the parts of the square depend on one another once the parameters are known: which families of rows and which families of columns are conditionally independent, and on what a single cell, or a block of cells in a group of rows, depends when the parameters, some rows and some columns are given.

Arrays and Their Laws

An array of order nn over a space YY, indexed by II, is a function yy from the index tuples v∈Inv\in I^n to YY; the arrays form the product YInY^{I^n}✦, and an array of order zero is identified with its single entry, a point of YY.✦ A map f:Y→Zf\colon Y\to Z acts on arrays coordinatewise, f(y)v=f(yv)f(y)_v=f(y_v).✦ For an array ww of order n+1n+1, an axis a∈{0,…,n}a\in\{0,\dots,n\} and an index i∈Ii\in I, the plane restriction w|a=iw|_{a=i} is the array of order nn whose entry at vv is the entry of ww at the tuple obtained by inserting ii into vv at position aa.✦ We write wi,⋅=w|0=iw_{i,\cdot}=w|_{0=i} and w⋅,i=w|n=iw_{\cdot,i}=w|_{n=i} for the planes through the first and the last axis, and for subsets RR and JJ of II we write wR,⋅=(wc,⋅)c∈Rw_{R,\cdot}=(w_{c,\cdot})_{c\in R} and w⋅,J=(w⋅,j)j∈Jw_{\cdot,J}=(w_{\cdot,j})_{j\in J}, the tuple of planes along JJ.

For random variables FF and GG under a probability measure μ\mu and a Markov kernel κ\kappa, we write ℒμ(F∣G)=κ(G)\mathcal L_\mu(F\mid G)=\kappa(G), or ℒμ(F∣G=g)=κ(g)\mathcal L_\mu(F\mid G=g)=\kappa(g), when κ\kappa is a version of the conditional law of FF given GG; a variable named ww is an array distributed by the measure in the subscript. For Markov kernels κi\kappa_i from SS to SiS_i, with ii in a finite set, the product kernel ⨂iκi\bigotimes_i\kappa_i sends ss to the product measure ⨂iκi(s)\bigotimes_i\kappa_i(s).✦

Fix a measurable automorphism TT of a standard Borel space XX. Initial array laws mXm^X over (X,T)(X,T) are probability measures mnXm^X_n on XInX^{I^n}, one for each nn, that are invariant under TT acting coordinatewise, consistent, every plane restriction carrying mn+1Xm^X_{n+1} to mnXm^X_n, and proper: for every proper subset JJ of II, the planes w⋅,Jw_{\cdot,J} are distributed under mn+1Xm^X_{n+1} as the product of |J||J| copies of mnXm^X_n.✦

An array object MM over initial array laws mXm^X consists of nonempty standard Borel spaces YY of states and UU of parameters with automorphisms TYT_Y and TUT_U, measurable maps x:Y→Xx\colon Y\to X and π:Y→U\pi\colon Y\to U with x∘TY=T∘xx\circ T_Y=T\circ x and π∘TY=TU∘π\pi\circ T_Y=T_U\circ\pi, and probability measures mnm_n on YInY^{I^n}, the laws of MM, subject to the following conditions.✦ Each mnm_n is invariant under TYT_Y and is carried by xx to mnXm^X_n; the laws are consistent and proper as above; under m0m_0, read on YY, the maps xx and π\pi are independent; and MM is conditionally local: for every nn and every proper subset JJ of II, with π(w)\pi(w) the parameter array of ww, obtained by applying π\pi coordinatewise, ℒmn+1(w⋅,J∣π(w))=⨂j∈Jλπ(w)⋅,j,\mathcal L_{m_{n+1}}\big(w_{\cdot,J}\mid\pi(w)\big)=\bigotimes_{j\in J}\lambda_{\pi(w)_{\cdot,j}},(1) where λp=ℒmn(w∣π(w)=p)\lambda_p=\mathcal L_{m_n}\big(w\mid\pi(w)=p\big) is the conditional law of an array of order nn given that its parameter array is pp. The Markov kernel p↦λpp\mapsto\lambda_p from UInU^{I^n} to YInY^{I^n} is the own kernel of MM.✦

Extensions and Saturation

An extension of an array object MM by an array object NN over the same initial array laws is a pair of measurable maps, a state map σ\sigma from the states of NN to those of MM and a parameter map ρ\rho from the parameters of NN to those of MM, that intertwine the automorphisms and satisfy xM∘σ=xNx_M\circ\sigma=x_N and πM∘σ=ρ∘πN\pi_M\circ\sigma=\rho\circ\pi_N, with σ\sigma, acting coordinatewise, carrying mnNm^N_n to mnMm^M_n for every nn.✦ The pullback along an extension (σ,ρ)(\sigma,\rho) of MM by NN at order nn is the linear isometry σ*\sigma^* from complex L2(mnM)L^2(m^M_n) to complex L2(mnN)L^2(m^N_n) given by F↦F∘σF\mapsto F\circ\sigma, with σ\sigma acting coordinatewise.✦

On its arrays of order n+1n+1, an array object MM has three kinds of data field, each a sub-σ\sigma-field of the measurable sets of YIn+1Y^{I^{n+1}}: the parameter field 𝒫n+1\mathcal P_{n+1}, and for each j∈Ij\in I the outside field 𝒪n+1,j\mathcal O_{n+1,j} and the enlarged field 𝒢n+1,j\mathcal G_{n+1,j}. The projection onto a data field 𝒟\mathcal D is the orthogonal projection of L2(mn+1)L^2(m_{n+1}) onto the 𝒟\mathcal D-measurable functions, the conditional expectation 𝔼M[⋅∣𝒟]\mathbb E_M[\,\cdot\mid\mathcal D].✦

An array object MM is saturated if its projections onto data fields commute with pullback along every extension: for every extension (N,σ,ρ)(N,\sigma,\rho) of MM, every nn, every kind of data field and every F∈L2(mn+1M)F\in L^2(m^M_{n+1}), 𝔼N[σ*F∣𝒟N]=σ*𝔼M[F∣𝒟M],\mathbb E_N\big[\sigma^*F\mid\mathcal D^N\big]=\sigma^*\,\mathbb E_M\big[F\mid\mathcal D^M\big],(2)✦ where 𝒟M\mathcal D^M and 𝒟N\mathcal D^N are the data fields of that kind of MM and of NN.

The Square

Let MM be an array object and yy an array of order n+2n+2 over its states, with parameter array E=π(y)E=\pi(y). Row aa of yy is the plane ya,⋅y_{a,\cdot} and column jj is the plane y⋅,jy_{\cdot,j}, both of order n+1n+1; the cell ya,j=(ya,⋅)⋅,jy_{a,j}=(y_{a,\cdot})_{\cdot,j} is an array of order nn. For sets AA and JJ of indices, yA,J=(ya,j)a∈A,j∈Jy_{A,J}=(y_{a,j})_{a\in A,\,j\in J} is the block of cells, with ya,Jy_{a,J} and yA,jy_{A,j} for the blocks of a single row and of a single column✦, and I\{j}I\setminus\{j\} is the set of indices outside jj✦; the rows and columns of EE are written in the same way. Conditional laws without a subscript are taken under mn+2m_{n+2}. A family of columns or of rows is proper when its index set is a proper subset of II.

For a set RR of rows, yR,⋅y_{R,\cdot} are the parent rows✦ and (E,yR,⋅)(E,y_{R,\cdot}) is the parent data. The parent entries of a column jj are the cells yR,jy_{R,j}. The raw data of a column jj is (E,y⋅,I\{j})(E,y_{\cdot,I\setminus\{j\}}), the parameters with the columns outside jj. The reveal data of jj for a set RR of rows is (E,y⋅,I\{j},yR,j)(E,y_{\cdot,I\setminus\{j\}},y_{R,j}), the raw data of jj together with the cells yR,jy_{R,j}. For sets PP and AA of rows and CC of columns, (E,yA,C)(E,y_{A,C}) is the group data and (E,yP,⋅,y⋅,C)(E,y_{P,\cdot},y_{\cdot,C}) the whole group data. The group selection sends the whole group data to the group data, the column restriction taking from the columns y⋅,Cy_{\cdot,C} their cells in the rows of AA, which form the block yA,Cy_{A,C}.✦

The row family kernel ΛA\Lambda_A of a set AA of rows is the Markov kernel sending a parameter array EE of order n+2n+2 to ΛA(E)=⨂a∈AλEa,⋅\Lambda_A(E)=\bigotimes_{a\in A}\lambda_{E_{a,\cdot}}.✦ The group marginal ΛA,J(E)\Lambda_{A,J}(E) of AA on a set JJ of columns is the image of ΛA(E)\Lambda_A(E) under the row restriction (wa)a∈A↦((wa)⋅,j)a∈A,j∈J(w_a)_{a\in A}\mapsto\big((w_a)_{\cdot,j}\big)_{a\in A,\,j\in J}.✦

For a Markov kernel κ\kappa from SS to ZZ and a measurable map φ\varphi on S×ZS\times Z, the observed kernel of κ\kappa and φ\varphi sends ss to the joint law of (φ(s,z),z)(\varphi(s,z),z) for zz drawn from κ(s)\kappa(s). For a parameter array pp of order n+1n+1 and a set RR of indices, the observed kernel of RR is the observed kernel of λ\lambda and φ(p,w)=wR,⋅\varphi(p,w)=w_{R,\cdot}; it sends pp to the joint law of (wR,⋅,w)(w_{R,\cdot},w) for ww drawn from λp\lambda_p.✦ The posterior λp(⋅∣wR,⋅=z)\lambda_p(\,\cdot\mid w_{R,\cdot}=z) of λp\lambda_p given the rows in RR, for arrays z=(zc)c∈Rz=(z_c)_{c\in R} of order nn, is the conditional law of ww drawn from λp\lambda_p given that its rows in RR are zz, a disintegration of the observed kernel of RR at pp.✦ The column kernel of a column jj is the Markov kernel sending the parent data (E,yR,⋅)(E,y_{R,\cdot}) to the posterior λE⋅,j(⋅∣wR,⋅=yR,j)\lambda_{E_{\cdot,j}}(\,\cdot\mid w_{R,\cdot}=y_{R,j}).✦

The old data of an array ww of order n+1n+1 at an index jj is (π(w),w⋅,I\{j})\big(\pi(w),w_{\cdot,I\setminus\{j\}}\big), and the row data map of a row cc sends the raw data (E,y⋅,I\{j})(E,y_{\cdot,I\setminus\{j\}}) to the old data (Ec,⋅,yc,I\{j})(E_{c,\cdot},y_{c,I\setminus\{j\}}) of row cc.✦ The row kernel of cc at jj sends the raw data to the conditional law under mn+1m_{n+1} of w⋅,jw_{\cdot,j} given the old data of ww, evaluated at the row data of cc.✦

The Claim

Theorem 1. Let MM be a saturated array object over initial array laws, with finite index set II, let n≥0n\ge0, let yy be an array of law mn+2m_{n+2} and let E=π(y)E=\pi(y). For every proper subset AA of II, ℒ(yA,⋅∣E)=⨂a∈AλEa,⋅.\mathcal L\big(y_{A,\cdot}\mid E\big)=\bigotimes_{a\in A}\lambda_{E_{a,\cdot}}.(3)✦ For every set RR of rows with |R|+2≤|I||R|+2\le|I| and every proper subset JJ of II, ℒ(y⋅,J∣E,yR,⋅)=⨂j∈JλE⋅,j(⋅∣wR,⋅=yR,j),\mathcal L\big(y_{\cdot,J}\mid E,y_{R,\cdot}\big) =\bigotimes_{j\in J}\lambda_{E_{\cdot,j}}\big(\,\cdot\mid w_{R,\cdot}=y_{R,j}\big),(4)✦ and, for n=0n=0, for every row a∉Ra\notin R and every column jj, ℒm2(ya,j∣E,y⋅,I\{j},yR,j)=ℒm1(w⋅,j∣π(w)=Ea,⋅,w⋅,I\{j}=ya,I\{j}).\mathcal L_{m_2}\big(y_{a,j}\mid E,y_{\cdot,I\setminus\{j\}},y_{R,j}\big) =\mathcal L_{m_1}\big(w_{\cdot,j}\mid\pi(w)=E_{a,\cdot},\, w_{\cdot,I\setminus\{j\}}=y_{a,I\setminus\{j\}}\big).(5)✦ If |I|≥2|I|\ge2, then for every row aa and every column j0j_0, ℒ(ya,⋅∣y⋅,I\{j0})=ℒmn+1(w∣w⋅,I\{j0}=ya,I\{j0})\mathcal L\big(y_{a,\cdot}\mid y_{\cdot,I\setminus\{j_0\}}\big) =\mathcal L_{m_{n+1}}\big(w\mid w_{\cdot,I\setminus\{j_0\}}=y_{a,I\setminus\{j_0\}}\big)(6)✦ and ℒ(y⋅,I\{j0}∣ya,⋅)=⨂j≠j0ℒmn+1(w∣wa,⋅=ya,j).\mathcal L\big(y_{\cdot,I\setminus\{j_0\}}\mid y_{a,\cdot}\big) =\bigotimes_{j\neq j_0}\mathcal L_{m_{n+1}}\big(w\mid w_{a,\cdot}=y_{a,j}\big).(7)✦ Let PP and AA be disjoint sets of rows with |P|+2≤|I||P|+2\le|I| and P∪A≠IP\cup A\neq I. For disjoint sets CC and JJ of columns with C∪J≠IC\cup J\neq I, ℒ(yA,J∣E,yP,⋅,y⋅,C)=ΛA,J(E).\mathcal L\big(y_{A,J}\mid E,y_{P,\cdot},y_{\cdot,C}\big)=\Lambda_{A,J}(E).(8)✦ If moreover |P|+|A|+1≤|I||P|+|A|+1\le|I|, then for every column jj and every split of I\{j}I\setminus\{j\} into disjoint sets CC and JJ, ℒ(yA,{j}∪J∣E,yP,⋅,y⋅,C)=ℒ(yA,{j}∪J∣E,yA,C),\mathcal L\big(y_{A,\{j\}\cup J}\mid E,y_{P,\cdot},y_{\cdot,C}\big) =\mathcal L\big(y_{A,\{j\}\cup J}\mid E,y_{A,C}\big),(9)✦ and for every split of II into disjoint sets CC and DD with DD nonempty, ℒ(yA,D∣E,yP,⋅,y⋅,C)=ℒ(yA,D∣E,yA,C).\mathcal L\big(y_{A,D}\mid E,y_{P,\cdot},y_{\cdot,C}\big) =\mathcal L\big(y_{A,D}\mid E,y_{A,C}\big).(10)✦

In the paper, the result reads: in m2m_2, simultaneously for EE in one conull set, (i) given EE, proper families of columns are independent with laws λE⋅,j\lambda_{E_{\cdot,j}} and proper families of rows are independent with laws λEa,⋅\lambda_{E_{a,\cdot}}, (ii) for |R|≤|I|−2|R|\le|I|-2, given EE and the parent rows yR,⋅y_{R,\cdot}, proper families of columns are independent, each with λE⋅,j\lambda_{E_{\cdot,j}} conditioned on its parent entries, and (iii) the law of ya,j0y_{a,j_0} given EE, yR,⋅y_{R,\cdot} and the other columns is λEa,⋅\lambda_{E_{a,\cdot}} conditioned on the rest of row aa, while in grouped form, for partitions of the rows into PP, AA, BB and of the columns into rr, ss, tt, the grouped coordinates are pairwise independent and ℒ(yA,t∣E,yP,⋅,y⋅,r∪s)=ΛA(dyA,t∣yA,r,yA,s)\mathcal L(y_{A,t}\mid E,y_{P,\cdot},y_{\cdot,r\cup s})=\Lambda_A(dy_{A,t}\mid y_{A,r},y_{A,s}) in every ordering, the paper writing kk for |I||I|, pap_a for Ea,⋅E_{a,\cdot} and qjq_j for E⋅,jE_{\cdot,j}. The library states these laws at every order n+2n+2, with cells that are arrays of order nn, except (5), stated at order two, and the column half of (i) is the conditional locality (1) of the object.

What Is Checked

The claim is stated over a saturated array object MM over initial array laws of a measurable automorphism TT of a standard Borel space XX, with finite index set II: over its laws mnm_n and own kernel λ\lambda, the planes of its arrays, and the row family kernels, group marginals, column kernels and row kernels built from them. The plane restrictions, the row data maps and the group selection are measurable✦.